AUDIO DENOISING AND NORMALIZATION USING IMAGE TRANSFORMING NEURAL NETWORK

Techniques are disclosed for reducing noise from an audio signal. A methodology implementing the techniques according to an embodiment includes generating a 2-dimensional (2D) spectrogram of a received audio signal and applying the 2D spectrogram to an image transformation neural network that is trained to transform the 2D spectrogram to generate an output spectrogram representing a denoised version of the received audio signal. The method further includes converting the output spectrogram to the time domain to generate the denoised audio signal. The neural network is trained on spectrogram images of clean and corrupted versions of training audio signals such that the trained neural network converts a spectrogram image of a corrupted audio signal into a spectrogram image more closely resembling a spectrogram of the associated clean audio signal. The denoising may also include removal of other degradation effects, including reverberation, unwanted voices, and unwanted music, from an audio signal.